Ensuring the unfold_backward op test runs on all backend

Move TORCH_CHECK_VALUE for step > 0 from CPU kernel to shared
unfold_backward function to ensure all backends validate the
parameter.

aten/src/ATen/native/UnfoldBackward.cpp

@@ -21,6 +21,7 @@ Tensor unfold_backward(
   int64_t size,
   int64_t step
 ) {
+  TORCH_CHECK_VALUE(step > 0, "step is ", step, " but must be > 0");
   auto grad_input = at::zeros(input_sizes, grad.options());
   if (step >= size) {
     auto gI_unfolded = grad_input.unfold(dim, size, step);

test/test_shape_ops.py

@@ -843,6 +843,15 @@ class TestShapeOps(TestCase):
         with self.assertRaisesRegex(RuntimeError, "step is -1 but must be > 0"):
             x.unfold(0, 1, -1)
 
+    def test_unfold_backward_errors(self, device):
+        grad_in = torch.randn(2, 3, device=device)
+        input_sizes = [6]
+
+        with self.assertRaisesRegex(ValueError, "step is 0 but must be > 0"):
+            torch.ops.aten.unfold_backward(grad_in, input_sizes, 0, 3, 0)
+
+        with self.assertRaisesRegex(RuntimeError, "size is -1 but must be >= 0"):
+            torch.ops.aten.unfold_backward(grad_in, input_sizes, 0, -1, 1)
 
 instantiate_device_type_tests(TestShapeOps, globals())